Agent Testing Patterns

Test AI agent systems that use tools, make decisions, and produce non-deterministic outputs.

Testing Challenges

Test Pyramid for Agents

        ╱╲
       ╱  ╲        E2E Agent Tests (few, expensive)
      ╱────╲       Full agent loop with real LLM
     ╱      ╲
    ╱────────╲     Integration Tests (moderate)
   ╱          ╲    Tool execution, state management
  ╱────────────╲
 ╱   Unit Tests ╲  Tool implementations, parsers, validators
╱────────────────╲

Unit Testing (No LLM)

Tool Implementation Tests

import pytest

def test_file_read_tool():
    tool = FileReadTool()
    result = tool.execute({"path": "test.txt"})
    assert result["content"] == "expected content"
    assert result["success"] is True

def test_file_read_tool_missing_file():
    tool = FileReadTool()
    result = tool.execute({"path": "nonexistent.txt"})
    assert result["success"] is False
    assert "not found" in result["error"].lower()

def test_tool_input_validation():
    tool = FileReadTool()
    with pytest.raises(ValueError, match="path is required"):
        tool.execute({})

Response Parser Tests

def test_parse_tool_call():
    raw = '{"tool": "search", "args": {"query": "python"}}'
    result = parse_tool_call(raw)
    assert result.tool == "search"
    assert result.args == {"query": "python"}

def test_parse_malformed_tool_call():
    raw = "not json at all"
    result = parse_tool_call(raw)
    assert result is None

Integration Testing (Mocked LLM)

Mock LLM Client

class MockLLMClient:
    def __init__(self, responses: list[dict]):
        self.responses = iter(responses)
        self.calls: list[dict] = []

    async def generate(self, messages: list[dict], tools: list[dict] = None) -> dict:
        self.calls.append({"messages": messages, "tools": tools})
        return next(self.responses)

@pytest.fixture
def mock_agent():
    client = MockLLMClient(responses=[
        {"content": None, "tool_calls": [{"name": "search", "args": {"query": "python packaging"}}]},
        {"content": "Based on the search results, here's how to package Python..."},
    ])
    return Agent(llm=client, tools=[SearchTool(), FileReadTool()])

Tool Execution Sequence Tests

@pytest.mark.asyncio
async def test_agent_uses_search_then_responds(mock_agent):
    result = await mock_agent.run("How do I package a Python project?")

    # Verify tool was called
    assert len(mock_agent.tool_history) == 1
    assert mock_agent.tool_history[0].tool_name == "search"
    assert "python" in mock_agent.tool_history[0].args["query"].lower()

    # Verify final response exists
    assert result.content is not None
    assert len(result.content) > 0

State Management Tests

@pytest.mark.asyncio
async def test_session_preserves_context(mock_agent):
    await mock_agent.run("My name is Alice")
    result = await mock_agent.run("What's my name?")

    # Verify conversation history maintained
    assert len(mock_agent.messages) == 4  # 2 user + 2 assistant

E2E Testing (Real LLM)

Structural Assertions

@pytest.mark.e2e
@pytest.mark.asyncio
async def test_agent_creates_file(real_agent, tmp_path):
    result = await real_agent.run(f"Create a Python hello world script at {tmp_path}/hello.py")

    # Assert on outcome, not exact content
    hello_file = tmp_path / "hello.py"
    assert hello_file.exists()
    content = hello_file.read_text()
    assert "print" in content  # Must use print
    assert content.strip()  # Non-empty

    # Verify it's valid Python
    compile(content, "hello.py", "exec")

Semantic Assertions

@pytest.mark.e2e
@pytest.mark.asyncio
async def test_agent_explains_concept(real_agent):
    result = await real_agent.run("Explain what a circuit breaker pattern is in 2-3 sentences")

    # Semantic checks (not exact string matching)
    assert len(result.content) > 50
    assert len(result.content) < 1000
    assert any(term in result.content.lower() for term in ["fault", "failure", "threshold", "open", "closed"])

Evaluation Metrics

@dataclass
class AgentEvalResult:
    task_completed: bool
    tool_calls_count: int
    tokens_used: int
    latency_ms: float
    error_recovery_count: int

async def evaluate_agent(agent, test_cases: list[dict]) -> list[AgentEvalResult]:
    results = []
    for case in test_cases:
        start = time.perf_counter()
        try:
            result = await agent.run(case["prompt"])
            completed = case["validator"](result)
        except Exception:
            completed = False
        latency = (time.perf_counter() - start) * 1000

        results.append(AgentEvalResult(
            task_completed=completed,
            tool_calls_count=len(agent.tool_history),
            tokens_used=agent.total_tokens,
            latency_ms=latency,
            error_recovery_count=agent.error_count,
        ))
    return results

Regression Testing

Golden File Tests

def test_tool_call_format_regression():
    """Ensure tool call format hasn't changed."""
    response = agent.format_tool_call("search", {"query": "test"})
    expected = load_golden("tool_call_format.json")
    assert response == expected

Benchmark Suite

BENCHMARK_CASES = [
    {"prompt": "List all Python files in the project", "expected_tools": ["glob"], "max_tokens": 500},
    {"prompt": "Fix the syntax error in app.py", "expected_tools": ["read", "edit"], "max_tokens": 2000},
]

async def run_benchmark(agent):
    for case in BENCHMARK_CASES:
        result = await agent.run(case["prompt"])
        tools_used = {t.tool_name for t in agent.tool_history}
        assert tools_used.issubset(set(case["expected_tools"] + ["think"]))
        assert agent.total_tokens <= case["max_tokens"]

Anti-Patterns

• Asserting exact LLM output — Models change; assert structure and semantics
• No mocking in unit tests — Real API calls make tests slow, expensive, and flaky
• Testing only happy path — Test error recovery, malformed responses, tool failures
• No cost tracking — Monitor token usage in E2E tests to catch regressions
• Ignoring non-determinism — Run E2E tests multiple times; set pass threshold (e.g., 4/5)
• Testing agent internals — Test outcomes and tool call patterns, not internal state